Paper 2604.08302
DMax: Aggressive Parallel Decoding for dLLMs
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 8
- GitHub
- 128 stars
01 In brief
Summary
DMax is a new paradigm for diffusion language models (dLLMs) that enables aggressive parallel decoding while preserving generation quality by mitigating error accumulation.
It reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings, using two key components: On-Policy Uniform Training (OPUT) and Soft Parallel Decoding (SPD).
OPUT extends a pretrained masked diffusion model by training it to recover clean tokens from both masked inputs and its own on-policy predictions, bridging the train-inference gap.
SPD represents intermediate decoding states as hybrid embeddings interpolated between predicted token embeddings and the mask embedding, providing uncertainty priors for self-correction.
Experiments on LLaDA-2.0-mini show significant improvements: on GSM8K, tokens per forward (TPF) increases from 2.04 to 5.48 with accuracy dropping only from 92.6% to 92.1%; on MBPP, TPF rises from 2.71 to 5.86 with comparable performance.
The model achieves over 1,000 tokens per second on two H200 GPUs.
DMax also improves accuracy at low parallelism, and ablations confirm that OPUT is a prerequisite for SPD.
02 From the paper
Abstract
We present DMax, a new paradigm for efficient diffusion language models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On two H200 GPUs, our model achieves an average of 1,338 TPS at batch size 1. Code is available at: https://github.com/czg1225/DMax